Patch Resume

Upload any resume → Claude turns it into clean, ATS-ready LaTeX → improve it by chatting or by pasting a job description → condense it to exactly one page → export.

The idea underneath it: a safe way to let an LLM edit a structured document. The model proposes a change, the system validates it against a hard check (it has to compile, it has to be one page), and a human approves — so nothing broken ever ships.

---

What it does

• 📤 Upload a PDF, DOCX, or TXT resume → text is extracted and Claude (claude-opus-4-8) rewrites it into clean, ATS-friendly LaTeX → live PDF preview.
• 💬 Chat to edit in plain English. Every AI edit is test-compiled before it's accepted; if the LaTeX fails to build, the compiler error is fed back to Claude and auto-repaired.
• 🎯 Tailor to a job description. Paste a JD and it rewrites the summary, bullets, and skills to match the role — truthfully (it won't invent skills or employers).
• 📃 Fit to 1 page. One click runs a real loop: compile → count PDF pages → if it's over one, ask Claude to condense (tighten wording, drop the weakest bullet, never merge into run-ons) → recompile → repeat until it's exactly one page.

The thread that ties it together: structured generation → hard validation → human-in-the-loop.

---

Tech stack

Layer	Stack
Backend	Python, FastAPI, Anthropic SDK (claude-opus-4-8)
Front-end	React + TypeScript + Vite + Tailwind / shadcn-ui
PDF	Tectonic (LaTeX → PDF), PyPDF2 (page counting)
Parsing	PyPDF2 (PDF), python-docx (DOCX)

State is kept in-memory with a JSON backup — no database, Redis, or Docker required to run it.

---

Quick start

Prerequisites

• Python 3.11+
• Node.js 18+
• Tectonic (the LaTeX engine) installed and on your PATH — https://tectonic-typesetting.github.io
• An Anthropic API key — https://console.anthropic.com

1. Clone + configure

git clone https://github.com/rohitsudhakar1/patch-resume.git
cd patch-resume
cp env.example .env
# open .env and set ANTHROPIC_API_KEY=sk-ant-...

2. Backend (terminal 1)

pip install -r requirements.txt
python -m uvicorn main:app --app-dir backend --port 8000

On Windows, prefix with set PYTHONUTF8=1 (or $env:PYTHONUTF8="1" in PowerShell) so the debug logs render.

3. Front-end (terminal 2)

npm install
npm run dev

Open http://localhost:8080. (Backend API + docs at http://localhost:8000/docs.)

---

How to use

1. Drag in a resume (or use the included test_resume.txt) → clean PDF renders on the right.
2. Type an edit in the chat — e.g. "tighten my summary and add a metric." The change is validated, then the PDF updates.
3. Tailor to a job description — click the button, paste a JD (try test_jd.txt), hit Tailor my resume.
4. Fit to 1 page — click the toolbar button; it loops until the resume is a single page.

---

How it works (the loops)

Edit / tailor — validate then auto-repair (backend/main.py):

Claude rewrites the LaTeX  →  test-compile it
   ├─ compiles?  → accept, update the project, re-render the PDF
   └─ fails?     → send the compiler error back to Claude → retry (up to 3x)

Fit to one page — a real agentic loop (/llm/fit-one-page):

compile → count pages (PyPDF2)
   ├─ pages == 1 → done
   └─ pages > 1  → ask Claude to condense → recompile → count again  (cap: 5 iterations)

Every AI call goes through a single helper, claude_complete() — the model's output is always treated as a proposal, never trusted blind.

---

Project structure

patch-resume/
├── backend/
│   ├── main.py              # FastAPI app: ingest, chat-edit, tailor, fit-one-page, compile
│   ├── config.py            # settings
│   └── services/
│       ├── template_service.py  # structured data → clean ATS LaTeX
│       └── compile_service.py   # LaTeX → PDF via Tectonic
├── src/                     # React front-end (ChatPanel, PDFViewer, Workspace, …)
├── test_resume.txt          # sample resume input
├── test_jd.txt              # sample job description
├── env.example              # copy to .env and add your key
└── requirements.txt

Key endpoints

Method	Endpoint	Purpose
POST	/ingest	Upload a resume → parse → LaTeX
POST	/llm/chat	Natural-language edit (validated)
POST	/llm/fit-one-page	Loop-condense to a single page
GET	/artifact/pdf/{id}	Compiled PDF
GET	/health	Health check

Notes

• The AI features require ANTHROPIC_API_KEY. Without it, ingest falls back to a template render but chat/tailor/fit are disabled.
• Image-only (scanned) PDFs need Tesseract for OCR; text-based PDFs work out of the box.

License

MIT — see LICENSE.